Team models: held-out predictions
Every point is one game, predicted before it was played: each week's games are predicted by models refit on
the season's earlier games only (modeling/evaluate_models.py, the same cohort as the
leaderboard).
The x axis is each model's implied margin: the league's points scale times the log-odds of its held-out home
win probability (NCAA 7.25, NBA and WNBA 8.75, G League and EuroLeague 12.75, EuroCup 11.25). Only the margin
and points families predict a native spread, and receipts evaluated before 2026-11-22 converted their spreads
at 11 points per log-odd, so on those seasons an NCAA margin model's implied margin is 7.25/11 of its spread
(fit slope near 1.3). Points open the game's box score where its box is stored, else that day's games.
End-of-season fits are never plotted here.
Model catalog
· Player models
· RAPM debugger.
Predicted (implied) vs actual margin
Home minus away; the dashed line is y = x, solid lines are each model's least-squares fit (slope under 1: the model's margins are too wide). Hover a point for the game; click it to open it.
Error over the season
Log loss on the winner (−ln of the winner's predicted probability), Brier, margin MAE and accuracy (a 50% pick counts half), averaged over each day's held-out games. Lower is better except accuracy.
Which seasons have held-out predictions
Plotted: seasons with a comparable full-season evaluation and its point set. The first three weeks of every season are training only (no predictions). Preseason priors are retrospective: when they were written is not recorded.
| League | Seasons plotted | Range | Games | Protocol | Not plotted |
|---|---|---|---|---|---|
| NCAAM | 38 | 1988-1989 to 2025-2026 | 154,630 | weekly refits on shared v2 matchup states, retrospective preseason priors; weekly refits, no preseason prior; weekly refits, retrospective preseason prior blend | 2026-2027 has not had enough games yet: results start once three weeks of games are in (the first holdout week is predicted from 21 days of earlier games) |
| NCAAW | 22 | 2004-2005 to 2025-2026 | 90,438 | weekly refits on shared v2 matchup states, retrospective preseason priors; weekly refits, no preseason prior; weekly refits, retrospective preseason prior blend | 2026-2027 has not had enough games yet: results start once three weeks of games are in (the first holdout week is predicted from 21 days of earlier games) |
| NCAAMD2 | 14 | 2012-2013 to 2025-2026 | 37,265 | weekly refits on shared v2 matchup states, retrospective preseason priors; weekly refits, no preseason prior; weekly refits, retrospective preseason prior blend | 2026-2027 has not had enough games yet: results start once three weeks of games are in (the first holdout week is predicted from 21 days of earlier games) |
| NCAAWD2 | 14 | 2012-2013 to 2025-2026 | 31,908 | weekly refits on shared v2 matchup states, retrospective preseason priors; weekly refits, no preseason prior; weekly refits, retrospective preseason prior blend | 2026-2027 has not had enough games yet: results start once three weeks of games are in (the first holdout week is predicted from 21 days of earlier games) |
| NCAAMD3 | 14 | 2012-2013 to 2025-2026 | 40,761 | weekly refits on shared v2 matchup states, retrospective preseason priors; weekly refits, no preseason prior; weekly refits, retrospective preseason prior blend | 2026-2027 has not had enough games yet: results start once three weeks of games are in (the first holdout week is predicted from 21 days of earlier games) |
| NCAAWD3 | 14 | 2012-2013 to 2025-2026 | 31,124 | weekly refits on shared v2 matchup states, retrospective preseason priors; weekly refits, no preseason prior; weekly refits, retrospective preseason prior blend | 2026-2027 has not had enough games yet: results start once three weeks of games are in (the first holdout week is predicted from 21 days of earlier games) |
| NBA | 31 | 1996-1997 to 2026-2027 | 32,833 | weekly refits on shared v2 matchup states, retrospective preseason priors; weekly refits, no preseason prior; weekly refits, retrospective preseason prior blend | - |
| WNBA | 30 | 1997 to 2026 | 5,449 | weekly refits on shared v2 matchup states, retrospective preseason priors; weekly refits, no preseason prior; weekly refits, retrospective preseason prior blend | - |
| GLEAGUE | 22 | 2004-2005 to 2025-2026 | 7,740 | weekly refits on shared v2 matchup states, retrospective preseason priors; weekly refits, no preseason prior; weekly refits, retrospective preseason prior blend | 2026-2027 has not had enough games yet: results start once three weeks of games are in (the first holdout week is predicted from 21 days of earlier games) |
| EUROLEAGUE | 20 | 2007-2008 to 2026-2027 | 4,305 | weekly refits on shared v2 matchup states, retrospective preseason priors; weekly refits, no preseason prior; weekly refits, retrospective preseason prior blend | - |
| EUROCUP | 20 | 2007-2008 to 2026-2027 | 2,932 | weekly refits on shared v2 matchup states, retrospective preseason priors; weekly refits, no preseason prior; weekly refits, retrospective preseason prior blend | - |